Essays on emerging markets

Texto completo

(1)ESSAYS ON EMERGING MARKETS. Xavier Ordeñana Rodríguez. TESI DOCTORAL UPF / 2015. DIRECTOR DE LA TESI. Dr. Jaume Ventura, CREI i Departament de Economia i Empresa DEPARTAMENT DE ECONOMIA I EMPRESA.

(2) ii.

(3) A Caty, Juan Xavier, Pablo José y Andrés.. iii.

(4)

(5) Acknowledgements I would like to thank my thesis supervisor, Jaume Ventura, for his guidance, patience and support throughout these years. The Dean of ESPAE Graduate School of Management, Virginia Lasio, has also been a key support in completing this dissertation. I want to thank special contributions of Gustavo Solórzano, coauthor of the paper for Chapter 3, Ramón Villa, coauthor of the paper for Chapter 1, participants at the Inter-American Development Bank Seminars, BALAS Annual Conference, International Breakfast Seminar at UPF as well as the Faculty Workshop at ESPAE. I appreciate research assistance by Ramón Villa and Andrea Samaniego. Finally, I would like to thank my family for their support during these years, specially my wife Caty and my three sons.. Xavier Ordeñana Universitat Pompeu Fabra. v.

(6)

(7) Abstract This thesis presents three essays on emerging Markets, with a special emphasis in Ecuador. The thesis uses both individual level (Chapter 1) and firm level data (Chapter 2) from Ecuador. Using dynamic pseudo panels, the first chapter finds an effect of entrepreneurship on the social mobility of Ecuadorians, with women experiencing a larger effect. The second chapter finds that FDI acts as external economies of scale, increasing the size of Ecuadorian firms. Finally, the third chapter presents a theoretical approach to the term maturity mismatch in Emerging Markets’ sovereign debt.. Resumen Esta tesis presenta tres ensayos sobre temas importantes para los países emergentes, con especial énfasis en Ecuador. La tesis utiliza tanto datos individuales (Capítulo 1) como de empresas (Capítulo 2), en ambos casos de Ecuador. Usando cuasi-paneles dinámicos, el primer capítulo encuentra un efecto del emprendimiento en la movilidad social de los ecuatorianos, con las mujeres experimentando un efecto mayor. En el segundo capítulo, se busca determinar los efectos de escala externa de la inversión extranjera sobre el tamaño de las empresas ecuatorianas encontrando un efecto positivo. Finalmente, el tercer capítulo presenta una aproximación teórica a la deuda de corto plazo que es utilizada por países emergentes como parte de su deuda soberana.. vii.

(8)

(9) Prologue This thesis presents three essays on emerging markets issues. It relates important topics such as entrepreneurship, foreign direct investment and debt. It is particularly focused in Ecuador, a middleincome country with significant opportunities and challenges in its way as a developing country. As many other developing countries, Ecuador is facing an “entrepreneurship boom”, with several public policies oriented towards entrepreneurial promotional. However, the actual impact ot these policies are not clear. The first chapter of the thesis aims to find an effect of entrepreneurship on income mobility, using constructed pseudo-panels of Ecuadorian individual data. Following other regional works, we observe a higher unconditional mobility than the regional average. Moreover, we find that indeed entrepreneurship acts as a catalyst for income mobility, with an even higher effect when considering only women entrepreneurs. The role of Foreign Direct Investment in emerging markets is also of great importance. In Chapter 2, we study the determinants of firm size using Ecuadorian firm data for the period 2007-2011. We particularly focus on the role of industry-level foreign direct investment on Ecuadorian firms’ size. We argue that FDI acts as an external economy of scale and thus generates spillover effect on. ix.

(10) other firms. We find a significant positive effect of industry-level FDI on firm size, when using assets as a measure for size. The final chapter switches to another recurrent emerging market issue: external debt. Financial crises throughout the years have showed the dangers of having high levels of short-term debt. We show that international investors prefer to finance long-term projects with short-term debt – that bears risk of self-fulfilling crises. Short-term debt enlarges their menu of assets allowing them to extract surplus from domestic entrepreneurs. The risk of a selffulfilling crisis acts as an entry barrier preventing other funds from flowing into the country. Lenders can assure this equilibrium through some coordination device e.g. a hedge fund. The three chapters of this thesis present different critical aspects of emerging markets, with a special empirical look to Ecuador. However, the challenges are shared by many other Latin American countries (or other developing nations from other regions).. x.

(11)

(12) Table of Contents. 1.. MOBILITY AND ENTREPRENEURSHIP IN ECUADOR: A DYNAMIC PSEUDO-PANEL APPROACH ...................................................................................................... 1 1.1. Introduction....................................................................... 1 1.2. Literature Review ............................................................ 5 1.3. Database Treatment and Documentation ................... 7 1.4. Income Mobility and Entrepreneurship ...................... 24 1.5. Results ............................................................................ 32 1.6. Conclusion ...................................................................... 39. 2.. EXTERNAL ECONOMIES OF SCALE: FOREIGN DIRECT INVESTMENT AND FIRM SIZE IN ECUADOR ........................ 44 2.1 Introduction......................................................................... 44 2.2 Literature Review .............................................................. 45 2.3 The Model: Determinants of Firm size ........................... 47 2.4 Results ................................................................................ 54 2.5 Concluding Remarks ........................................................ 66. 3.. SHORT TERM BORROWING: EMERGING MARKETS’ ONLY CHOICE? ............................................................................................. 69 3.1. Introduction..................................................................... 69 3.2. A Model of Optimal Debt Structure in a Competitive Bond Market ............................................................................. 74 3.3. Equilibria and Welfare Comparisons .......................... 88 3.4. Departing from the Competitive Bond Market .......... 93 3.5. Concluding Remarks .................................................... 99. REFERENCES ....................................................................................................... 102 xii.

(13) xiii.

(14) 1. MOBILITY. AND. ECUADOR:. A. ENTREPRENEURSHIP DYNAMIC. IN. PSEUDO-PANEL. APPROACH1 1.1. Introduction There seems to be a consensus among policy makers in Latin America that promoting entrepreneurship is a way to achieve economic development. Different programs around the region, such as Emprende Ecuador and Start-Up Chile, exemplify this idea. However,. the. economic. effect. of. policies. that. promote. entrepreneurship at the country level is still unclear. In countries like Ecuador, where this study is focused, about one in five people is engaged in entrepreneurial activities, according to the Global Entrepreneurship Monitor Ecuador Report 2010. However, most of the entrepreneurial activity is highly ineffective at creating jobs. Ninety-eight percent of the entrepreneurs created fewer than five jobs. Shane (2009) suggests that these activities are not contributing to economic growth and thus should not be promoted by the government. However, Amorós and Cristi (2010) find a. 1. This chapter is part of the research project “Strengthening Mobility and Entrepreneurship: A Case for The Middle Classes” financed by the Inter-American Development Bank (IDB) Research Department. It was published in Latin American Journal of Economics (November 2014): http://www.lajece.org/docs/107764_laje_512307.pdf.

(15) positive effect on poverty reduction, which remains an important issue in Latin America, particularly in Ecuador. Across the world, entrepreneurship is becoming an attractive career option (see for example OECD (2012)). In the particular case of Ecuador, the percentage of people that label themselves as entrepreneurs has been consistently increasing in the last years (see Lasio et al, 2014). However, how successful is entrepreneurship in promoting social mobility? This chapter focuses on this important economic question: what is the effect of entrepreneurship on social mobility? Is there evidence that entrepreneurship increases a person’s relative income? In order to determine whether such a correlation exists, we studied the evolution of household income over time, using panel data. Unfortunately, in Ecuador, attempts to build rotating panels have only recently begun to be undertaken, and we encountered several problems when attempting to construct a database using this information. We found many statistical inconsistencies, and only short time spans are available for the construction of the data series. Techniques have been developed to remedy these limitations, and several authors have established that panel data are not necessary for many commonly estimated dynamic models (Heckman and Rob, 1985; Deaton, 1985; and Moffitt, 1990). To overcome the shortcomings created by the lack of panel data, a pseudo-panel approach was taken in order to estimate the dynamics of the model. This method was first introduced by Deaton (1985). 2.

(16) and consists of categorizing “similar” individuals in a number of cohorts, which can be constructed over time, and then treating the average values of the variables in the cohort as synthetic observations in a pseudo-panel. The dynamic nature of the model presented results in a series of complications when estimating the autoregressive parameters, as it introduces a missing variable problem that would render standard estimation methods biased. But the fact that the model is estimated at a cohort level provides an alternative to conventional dynamic panel estimation methods, as it can be assumed that the intrinsic cohort average effect will be asymptotically equal to 0 and, with large cohort dimensions, the missing variable problem is solved. Nevertheless this assumption raises serious questions about the validity of the constructed cohorts as units of study. If one is willing to assume that the intrinsic cohort effect is evened out, then how many idiosyncratic characteristics of the group in question could suffer the same attrition? Cohorts should be constructed on the basis of homogeneity of the individuals within the group, so as to avoid the loss of essential information and to make the mean estimators more significant. If it is assumed that this is not so, then even if the estimations are consistent, the conclusions derived from the study will not be very relevant. On account of this, we propose that the existence of idiosyncratic cohort effects should be taken as an indicator of accurately constructed cohorts. Thus, estimating by. 3.

(17) OLS the proposed model without accounting for the unobserved cohort effect will result in positively biased estimators. An unbiased estimator of the relevant model parameters is obtained through the use of traditional dynamic panel estimation methods (like the two-step least squares estimation or a more efficient estimator obtained by GMM methods).. The fact that OLS is. expected to provide a positive bias in the estimator if the idiosyncratic component exists is used as a useful check on the validity of the cohorts studied. We first calculate an unconditional convergence model and then we condition on entrepreneurship2. We find an unconditional convergence of 0.86 – slightly below the 0.90 of the region (Cuesta et al 2011). When entrepreneurship is included, the convergence decreases to 0.77, suggesting a relevant effect of entrepreneurship on income mobility. Moreover, as shown in the Results section, we find a higher mobility effect on the female cohorts suggesting a stronger relevance of entrepreneurship on women regarding social mobility. The rest of the chapter is organized as follows: Section 2 reviews related literature. Section 3 presents the data treatment and the construction of the pseudo-panel. Section 4 presents the models of mobility:. unconditional. and. conditional.. The. econometric. 2. We refer to entrepreneurs as those who declare to “own a business”, following Ecuadorian data. More information on Section 3. 4.

(18) techniques used for the estimation of the modeled and the assumptions taken are also discussed in this section. Section 5 presents the results and the analysis, and Section 6 concludes.. 1.2. Literature Review The study of income distribution on the basis of equality is one of the most important topics in the field of welfare economics. Traditionally a static approach was taken to provide a measurement of income inequality and welfare in any given society, but this method has a number of shortcomings as in most cases the study of social welfare requires the analysis of income dynamics as well. With the increased availability of panel and time series data, the primary focus of recent research in this area of science has centered more on income mobility. A more comprehensive overview of the basic highlights of the theory of income mobility are presented in Fields and Ok (1996) and Fields (2008). Due to the lack of extensive panel data in the region, only recently has the study of income mobility and its determinants been undertaken. This was primarily motivated by the development and increasing popularity of econometric techniques to remedy the lack of panel data. Cuesta et al (2011) use a pseudo-panel approach to study the differences in mobility across the Latin American region. They find a high level of unconditional mobility in the region and significant differences across countries. They also find that the measure of unconditional income mobility tends to underestimate. 5.

(19) the true mobility experienced by households in an economy. Canelas (2010) also uses this technique to measure poverty, inequality, and mobility in Ecuador and finds a decrease in poverty but persistent inequality between 2000 and 2009. With. the. recent. development. of. initiatives. to. compare. entrepreneurship across countries, like the Global Entrepreneurship Monitor or the World Bank Enterprise Data/Survey, research in this area has begun to study the effect that entrepreneurship has on income mobility and to what extent this translates on overall economic growth. Carree and Thurik (2005) provide a survey on the literature. linking. entrepreneurship. with. economic. growth.. Regarding the effect that entrepreneurship has on income mobility, Quadrini (2005), with the use of the Panel Study of Income Dynamics (which is a national survey conducted annually on a sample of U.S. families since 1968), finds that entrepreneurs experience greater upward mobility as they have a greater probability of moving to higher wealth classes and that this is not only a consequence of their higher incomes. But this does not hold for all types of entrepreneurship. For example Shane (2009), using data from several developed countries, shows that promoting a large number of start-ups is not a good public policy, since they do not create many jobs or significantly contribute to economic growth. The most common view found in literature is that there is a particular entrepreneurship relevant in improving economic performance, and it is generally identified as high-growth startup business.. 6.

(20) As Lederman et al (2014) comment, policy makers should care about entrepreneurship because is a fundamental driver of growth and. development.. The. relative. size. of. necessity. based. entrepreneurship in Latin-American countries makes it very important to evaluate and determine the overall effect that public policy oriented in encouraging entrepreneurship will have on the equality and welfare of a society. However, the amount of literature discussing its effect on income mobility is very limited. Some authors, such as Kantis et al (2002) show the relevance of entrepreneurship in Asian social mobility (compared to Latin America) using descriptive statistics.. This. chapter fits in the literature by aiming to determine the magnitude of entrepreneurial effect on income mobility.. 1.3. Database Treatment and Documentation The main objective of this chapter is to describe the linkages between entrepreneurship and mobility. In this section, we first analyze the intragenerational mobility experienced by Ecuadorians (unconditional mobility) and approach the potential role of entrepreneurship in improving mobility (conditional mobility). Given that individual data panels are nonexistent in Ecuador, the use of pseudo-panels was required. We start by explaining how we constructed the instrument.. 7.

(21) a) Database Treatment The data used for the construction and estimation of the pseudopanel were obtained from the National Employment and Unemployment Survey (ENEMDU for its Spanish acronym) collected by the National Institute of Statistics and Census (INEC). The survey is taken at the national level each November, and the results are processed and made public the following month. This data collection methodology has been applied since 2003, as before this year only the urban population was sampled. In some years, national census data are presented in May or June. To avoid any seasonal bias due to variations in the levels of economic activity at different times of the year, only those surveys made public in December were used. The database used to estimate the pseudo-panel is constructed as a series of independent cross-sections, one for each period analyzed. To determine the period in which the pseudo panel ought to be constructed, it is first necessary to examine the changes made by INEC in the methodology for both the determination of the sample and the estimation of the relevant variables. There are two important changes made in the last decade in the ENEMDU’s methodology which are so significant that, without taking them in account, any estimation made for the whole period would suffer from serious bias. The first occurred in December of 2003, before which time only the urban population was analyzed. The definition of what constitutes an urban settlement was also changed in 2003 to include centers with more than 2,000 inhabitants rather than the 8.

(22) 5,000 used earlier. The definitions of several labor variables were also modified while others were included. The second set of changes introduced by INEC in September 2007 consisted of several modifications in labor market definitions and classifications, but there were no significant changes in the variables used for this study (even though income estimation suffered some alterations, which will be discussed below). Because of the loss of information that would result from the construction of a larger panel (in the time dimension), only the 2003-2010 period was studied. To maintain the consistency of the data for the period analyzed, special attention was paid to the changes in the methodology, variable classification, and labels used. The method used for the estimation of individual income was also changed, and new income criteria were introduced after 2007. In addition, even if a survey question was not modified, some of the responses were changed, which in some cases made it impossible to use the variable for the whole period. To account for all of these issues, the income series was constructed using the previous methodology, and all of the other variables included were previously processed to ensure their statistical comparability. Using this methodology, income is calculated as any payment, either monetary or in-kind, received by the individual on a regular basis (daily, weekly, or monthly). Only one type of income source was considered: income generated by work (either from a primary or secondary occupation). A monthly income series was then. 9.

(23) constructed by adding all income originating from this source of personal revenue. The ENEMDUs were processed in order to obtain the pertinent variables at household level, as the information relevant for this study on an individual level is collected by INEC. Data mining techniques were used and, with the use of Structured Query Language (SQL), income and other covariates were aggregated at the desired level. The first key concept in determining the effect of entrepreneurship on income mobility is the definition of which households are to be considered entrepreneurs. The focus of the study is those households. that. are. entrepreneurs. by. choice. rather. than. entrepreneurs due to lack of options (a group that is difficult to correctly identify considering the scant information available). In order to reduce the probability of error at the moment of classification, only those households in which at least one member currently employs other workers are considered entrepreneurs.. b) Construction of the Pseudo-panel In order to analyze the dynamic nature of income mobility, household income needs to be observed over time. Given the absence of panel data, a pseudo-panel must be constructed. The pseudo-panel. approach. consists. of. categorizing. “similar”. individuals into a number of cohorts, which can be constructed over time, and then treating the average values of the variables in the 10.

(24) cohort as synthetic observations in a pseudo-panel. Even though this approach has many limitations compared to real panel data, it reduces several problems characteristic of real panel data. First, it greatly diminishes the problem of sample attrition, hence allowing the possibility for the construction of larger panels in the time dimension. A second contribution is that, as the observations are obtained by averaging different observations in a cohort, the possibility of measurement error is greatly reduced (provided the cohorts are adequately constructed). The efficiency and consistency of the estimators depends, among other things, on the criterion used for the construction of the different cohorts and the asymptotic nature of the data assumed. Several of the requirements for the consistency of pseudo-panel estimation are discussed by Verbeek and Nijman (1992), who recommend that the choice of the variables for the discrimination of the cohorts in the sample should follow three criteria: •. The cohorts are chosen such that the unconditional probability of being in a particular cohort is the same for all cohorts.. •. The variables chosen should be constant over time for each individual, because individuals cannot move from one cohort to another. This maintains the independence of the different cohort observations.. •. These variables should be observed for all individuals in the sample. This could be remedied by the use of unbalanced. 11.

(25) panel methods, but due to the short time span of the constructed pseudo-panel, this alternative is not considered.. Following these assumptions, cohorts were constructed using gender and date of birth of the household head. To determine the number of cohorts to be constructed, first the distribution of the date of birth variable was tested with conventional goodness of fit methods, but no traditional distribution seemed to adjust the data correctly. To ensure relatively similar probabilities of belonging to a birth cohort, the aggregated data for year of birth for the eight periods were divided into deciles. This avoids the possibility that a cohort in a given period becomes too small to provide an accurate estimation of its true characteristics. After considering the weights of the observations due to sample stratification, the following deciles were obtained: Table 1. Date of Birth ( ) cohorts criterion 1934. 1934. 1942. 1954. 1958. 1942 1949 1958 1962 1966 1971. 1949 1954 1962 1966 1971 1977 12.

(26) 1977. Source: Authors’ calculations.. . Another criterion used to determine the number of cohorts is the gender of the household head ( ). The conjunction of the two. variables (considering the criterion proposed for the date of birth) results in 20 cohorts per year and 160 synthetic observations in the. pseudo panel. The distribution of the observations and their corresponding expanded population values (by the use of sampling weights) in each of the categories explained are presented in the appendix. The inclusion of the gender of the household head as a determinant for the conformation of the cohorts makes the probability of belonging to a cohort uneven for most cohorts (as male household heads are more frequently found). As the synthetic observations are calculated with different sample sizes, a systematic heteroskedasticity component is introduced to the error. The methods to correct this problem are discussed in Gurgand, Gardes, and Bolduc et al. (1997). This problem becomes less relevant in cohorts constructed with a large number of observations, since the variance of the mean approaches zero as this number tends to infinity.. c) Treatment of Outliers The household income series each year is irregular, as its standard deviation is between 2 and 4 times the mean. The asymmetries presented by the data may complicate the estimation of any inference model applied. This is also maintained at a cohort level, 13.

(27) and important differences in variances between each cohort average are observed. These differences make the heteroskedasticity component, described in the previous section, more important. To account for this problem, data mining techniques are applied to determine and exclude outliers. A median of absolute deviations (MAD) approach is used to determine outliers in each cohort as, due to the nature of the series, the median is a better central tendency measure than the mean. Under this scheme, the following univariant filter was applied to each observation, and observations that satisfy this restriction are considered outliers:. 103 . . As shown in the formula, the method is applied at a cohort level for each period. Approximately 1.2 percent of the sample was determined to be an outlier. In Figures 1 and 2, the average income estimated for male and female cohorts is presented before and after the MAD treatment is applied. The red dotted bands denote the 95 percent confidence interval for the estimated means (solid blue line) for the period analyzed. The graph to the left of each vertical black line corresponds to the estimated cohort mean of household incomes before the univariant filter is applied, and to the right the results excluding outliers are presented. In observing the two figures, it is noteworthy that the error bands on male cohorts are smaller than those observed for female cohorts, 3. Traditionally the tolerance criterion is set at 4.5, but this resulted in the loss of 12 percent of the sample, including an important percentage of entrepreneurs.. 14.

(28) possibly due to the lower number of observations in the female cohorts. But even after considering those wider confidence intervals, for most of the years studied and most of the birth cohorts, male-headed households experience a significantly higher income than female-headed households (at a 95 percent confidence level). Without the application of the MAD univariant filter, some birth cohorts exhibit very irregular behavior, and in the case of female cohorts some of the error bands explode (raising serious concerns about the validity of those estimations in a pseudo-panel context). However, once outliers are excluded, the error bands decrease considerably and the behavior of the income mean becomes smoother.. 15.

(29) Figure 1. Average Income Before and After MAD Treatment for Male Cohorts. 16.

(30) Figure 1 (cont). Average Income Before and After MAD Treatment for Male Cohorts. Source: Authors’ calculations.. 17.

(31) Figure 2. Average Income Before and After MAD treatment for Female Cohorts. 18.

(32) Figure 2 (cont). Average Income Before and After MAD treatment for Female Cohorts. Source: Authors’ calculations.. 19.

(33) Additional descriptive statistics regarding the cohorts constructed after the outlier treatment are presented next. For most of the male cohorts, the percentage of households located in urban areas is significantly lower than the one presented by female cohorts. This might be due to a more traditional family life in rural areas, which makes single-parent families, or female-headed households, a less common occurrence. It is also important to note that the number of entrepreneur households headed by women is significantly lower than those headed by men. This is accentuated by the fact that almost half of all female entrepreneur households include a member, different from the household head, who owns a business.. 20.

(34) Table 2. Urban Ratio Male Households 1934 . 1934 1942 1942 1949 1949 1954 1954 1958 1958 1962 1962 1966 1966 1971 1971 1977 1977. . Female Households. 200 3. 200 4. 200 5. 200 6. 200 7. 200 8. 200 9. 201 0. 200 3. 200 4. 200 5. 200 6. 200 7. 200 8. 200 9. 201 0. 56%. 58%. 56%. 54%. 53%. 53%. 53%. 54%. 59%. 63%. 64%. 66%. 62%. 63%. 62%. 63%. 59%. 61%. 55%. 56%. 61%. 60%. 60%. 59%. 68%. 67%. 62%. 63%. 65%. 67%. 66%. 71%. 64%. 64%. 62%. 61%. 61%. 61%. 59%. 58%. 71%. 74%. 72%. 73%. 71%. 72%. 72%. 72%. 66%. 67%. 65%. 66%. 66%. 68%. 63%. 63%. 75%. 75%. 79%. 79%. 77%. 77%. 74%. 76%. 70%. 70%. 67%. 68%. 67%. 68%. 67%. 66%. 81%. 81%. 80%. 77%. 80%. 75%. 74%. 73%. 69%. 68%. 70%. 71%. 67%. 69%. 63%. 66%. 78%. 80%. 78%. 76%. 83%. 84%. 78%. 79%. 68%. 69%. 68%. 73%. 68%. 67%. 67%. 66%. 72%. 76%. 78%. 79%. 78%. 76%. 77%. 77%. 66%. 67%. 68%. 66%. 67%. 66%. 66%. 66%. 77%. 76%. 78%. 82%. 79%. 78%. 81%. 80%. 67%. 69%. 70%. 69%. 68%. 67%. 70%. 68%. 76%. 73%. 80%. 74%. 77%. 71%. 82%. 82%. 71%. 69%. 71%. 75%. 70%. 70%. 73%. 73%. 84%. 84%. 86%. 77%. 84%. 82%. 86%. 83%. Source: Authors’ calculations.. Table 3. Entrepreneurship Ratio. 21.

(35) Male Households 1934 . 1934 1942 1942 1949 1949 1954 1954 1958 1958 1962 1962 1966 1966 1971 1971 1977 1977. . Female Households. 200 3. 200 4. 200 5. 200 6. 200 7. 200 8. 200 9. 201 0. 200 3. 200 4. 200 5. 200 6. 200 7. 200 8. 200 9. 201 0. 6%. 11%. 9%. 9%. 5%. 6%. 4%. 3%. 3%. 3%. 2%. 4%. 2%. 3%. 2%. 2%. 9%. 13%. 11%. 11%. 8%. 9%. 7%. 5%. 5%. 8%. 5%. 7%. 5%. 5%. 3%. 2%. 9%. 13%. 13%. 13%. 11%. 13%. 7%. 7%. 4%. 10%. 5%. 6%. 5%. 5%. 3%. 4%. 9%. 12%. 12%. 14%. 12%. 11%. 9%. 7%. 7%. 7%. 7%. 10%. 6%. 6%. 4%. 3%. 9%. 15%. 12%. 16%. 8%. 11%. 10%. 8%. 7%. 7%. 5%. 7%. 6%. 5%. 6%. 3%. 8%. 13%. 12%. 12%. 12%. 10%. 10%. 7%. 5%. 4%. 5%. 5%. 3%. 5%. 5%. 4%. 8%. 12%. 10%. 13%. 10%. 11%. 7%. 8%. 3%. 7%. 6%. 6%. 5%. 6%. 3%. 1%. 5%. 10%. 10%. 10%. 10%. 10%. 8%. 9%. 5%. 6%. 4%. 6%. 6%. 3%. 2%. 3%. 6%. 9%. 8%. 9%. 9%. 8%. 6%. 5%. 3%. 5%. 4%. 7%. 3%. 4%. 1%. 3%. 5%. 6%. 5%. 6%. 3%. 4%. 4%. 2%. 3%. 3%. 2%. 3%. 2%. 5%. 1%. 1%. Source: Authors’ calculations.. Table 4. Percentage of Entrepreneurs who are Household Heads. 22.

(36) Male-headed Households 1934 1934 1942 1942 1949 1949 1954 1954 1958 1958 1962 1962 1966 1966 1971 1971 1977 1977. . Female-headed Households. 2003. 2004. 2005. 2006. 2007. 2008. 2009. 2010. 2003. 2004. 2005. 2006. 2007. 2008. 2009. 2010. 78%. 69%. 76%. 71%. 72%. 76%. 62%. 68%. 37%. 52%. 30%. 49%. 58%. 40%. 21%. 42%. 76%. 88%. 88%. 87%. 72%. 77%. 70%. 77%. 26%. 35%. 34%. 52%. 58%. 45%. 49%. 39%. 73%. 82%. 85%. 79%. 80%. 80%. 82%. 69%. 53%. 66%. 70%. 50%. 63%. 80%. 56%. 59%. 78%. 81%. 78%. 82%. 86%. 79%. 87%. 80%. 84%. 74%. 75%. 66%. 79%. 44%. 39%. 40%. 77%. 80%. 80%. 78%. 82%. 89%. 82%. 90%. 61%. 81%. 57%. 58%. 84%. 80%. 81%. 66%. 85%. 82%. 90%. 81%. 88%. 84%. 78%. 81%. 61%. 84%. 72%. 85%. 74%. 84%. 88%. 77%. 90%. 84%. 83%. 82%. 80%. 92%. 94%. 92%. 55%. 79%. 82%. 96%. 71%. 90%. 95%. 87%. 85%. 79%. 89%. 85%. 88%. 93%. 89%. 86%. 66%. 67%. 60%. 86%. 70%. 96%. 83%. 97%. 84%. 86%. 83%. 87%. 86%. 86%. 86%. 82%. 95%. 93%. 75%. 87%. 100 %. 82%. 87%. 90%. 79%. 91%. 79%. 98%. 90%. 77%. 82%. 90%. 86%. 60%. 100 %. 70%. 79%. 62%. 74%. 87%. Source: Authors’ calculations.. 23.

(37) 1.4. Income Mobility and Entrepreneurship a) The Unconditional Model Income mobility presents a measure of the relationship between past and present income. This relationship can be represented by the following equation:. !, # $!,% & ',. Where !, represents the household income on period (, ', a. composite error term and $ is a measure of the unconditional. income convergence ($=0 represents a perfect income mobility and. $=1 represents an absence of income mobility or perfect. convergence).. Since the information for the same individual is not available in the different years sampled, a pseudo panel approach was taken in order. to estimate the $ parameter. As previously mentioned the synthetic. observations are constructed with the average values of the household observations in each cohort. The dependent variable used for the estimation of the model is the log of the average household’s income for the cohort and the period studied, which makes the $. parameter a measure of the elasticity of past and present income. The respective cohort model can be expressed as follows: ln ,!-, . # $ ln ,!-,% . & '-,. 24.

(38) For ease of exposition, the logarithm of the income variable for each cohort will still be represented as !-, .. Fields and Ok (1999) demonstrate that this measure of income mobility is the only one to have a set of desired properties (scale invariance, symmetry, multiplicability, and additive separability).. b) The Conditional-Entrepreneurship Model As shown in Cuesta and Ñopo (2011), the measure of unconditional income mobility tends to underestimate the true mobility experienced by households in an economy. The effect of household covariates on income mobility can be estimated by an extension of !-, # $ !-,% & $ /0,% !-,% & $1 /0, & $2 3 !-,% & $4 ---5 ,.,. the previous model.. Where:. & $6 ---5 ,. & '-,. !-, = Represents the Neperian logarithm of the average. household income of cohort 7 in period (. The income. variable is previously deflated considering the purchasing power parity (PPP) index reported by the World Bank, thus allowing inter-country comparison.. /0, = Is the proportion of households which are considered. to be entrepreneurs in cohort 7 and period (. This regressor. is believed to be predetermined, a concept that will be clarified below.. 25.

(39) 3 = Is a dichotomic variable which takes the value of 1 if 7. ---- ,., = A set of time-variant household covariate averages 5 is a female cohort and 0 otherwise.. for cohort 7 and period (. These regressors are believed to be. ---5 ,. = A set of time-invariant household covariate averages exogenous.. for cohort 7. The term 7,(. is included to denote that the average is taken on period ( and, as the sample mean is an error-driven measurement, differences may be observed in. time. These variables need not be truly time-invariant, but the rate at which they vary may be too subtle to be observed in one period (variables that present a staircase behavior and which require more than one period to register a change fall in this category). An obvious example is the gender of the household head, due to the construction of the cohorts. No assumptions about the relationship of these covariates with the error term are made.. $4 and $6= Are vectors of the pseudo-elasticity of said. '-, = A composite error term determined by the next covariates on present incomes.. equation.. '-, # 90 & :-,. Where 90 is a time-invariant intrinsic cohort “7” component which cannot be observed, and :-, is an error term. The different possible. assumptions for these terms are considered below.. 26.

(40) ;!-, # $ & $ /0,% & $2 3 ;!-,%. The total measure of income mobility can be expressed as:. Where. $= Represents the part of income convergence that is only explained by past income.. $= is the effect that a marginal increase in the. entrepreneurship percentage in the cohort has on its income convergence.. $2 = represents the variation in income convergence. experienced by female cohorts.. ---- ,. will provide accurate estimators of the ---- ,., and 5 !-, , /0, , 5 If the number of observations in each cohort is sufficiently large,. true cohort means. An unbiased estimator of the relevant model. parameters can be obtained through traditional dynamic panel estimation methods (like the two-step least squares estimation or a more efficient estimator obtained by GMM methods).. c) Estimation by Dynamic Panel Methods The dynamic nature of the model presented results in a series of complications when estimating the autoregressive parameters. Because the estimation centers on following the same cohorts over time, as previously indicated, an unobserved fixed component is introduced into the equation. The following expression is obtained by replacing the composite error term with its determinants:. 27.

(41) !-, # $ !-,% & $ /0,% !-,% & $1 /0, & $2 3 !-,% & $4 ---5 ,., & $6 ---5 ,. & 90 & :-, ,1.. By expressing this equation on the t 1 period, one can show that. the unobserved component is correlated with the previous income (this is also likely to hold for the other contemporaneous covariates), which introduces a missing variable problem that would render standard estimation methods biased. However, several assumptions can be made about the nature of the error components (depending on the licenses the researcher is willing to take), and from which different consistent estimation methods can be derived. The following assumptions are made about the cohort covariates and the error components: •. The error term is believed to be serially uncorrelated,4. •. Entrepreneurship is believed to be predetermined,. =>?0, ?0,% @ # 0. =,/0,%A :-, . # 0. ,2.. B C0. ,3.. This means that the error term may be correlated with contemporaneous. or. future. entrepreneurship. levels.. This. assumption is made because entrepreneurship is believed to be. 4. This hypothesis can be tested and if the error’s autocorrelation cannot be rejected at a given confidence level, then other considerations, which are later specified, must be taken into account.. 28.

(42) endogenous, as it is difficult to establish a causal relationship between this variable and the contemporaneous average household income for each cohort. This assumption also serves as a source of instruments to account for the endogeneity problem derived from the described nature of entrepreneurship. •. The time-variant cohort covariates are assumed to be exogenous, but not strictly so.. •. ---- ,.,%A :-, @ # 0 =>5. B CD0. ,4.. No assumptions are made about the nature of the timeinvariant covariates.. •. Building on the belief that the errors are serially uncorrelated, a natural supposition is made: E,?0 , !-,% . # 0. ,5.. The fact that the model is estimated at a cohort level provides an alternative to conventional dynamic panel estimation methods. Moreover, if it is assumed that:. 9 ~G,0, H . B I 7 J B 7 I K. ,6.. then the intrinsic cohort average effect will be asymptotically equal to zero and, with large cohort dimensions, the missing variable problem is solved. Next, instrumental variables are needed to account for the endogeneity of entrepreneurship, after which a twostage OLS estimation method will provide consistent estimators (even though a robust estimation is recommended because of the existence of an important heteroskedastic factor caused by the use. 29.

(43) of sample averages as observations). But assumption (6) implies that no intrinsic cohort effect exists, which raises serious questions about the validity of the constructed cohorts as units of study. If one is willing to assume that the intrinsic cohort effect is evened out, then how many important effects suffer the same attrition? Cohorts should be constructed on the basis of homogeneity of the individuals within the group, so as to avoid the loss of essential information and to make mean estimators more significant. If it is assumed that this is not so, then even if the estimations are consistent, the conclusions derived from the study will not be very relevant. Hence, the existence of idiosyncratic cohort effects should be taken as an indicator of accurately constructed cohorts.. On. account of these issues, and the belief that the cohorts defined for this study are correctly specified, assumption (6) is relaxed. Thus, estimating the proposed model without accounting for the unobserved cohort effect will result in biased estimators and ordinary least squares (OLS) methods will result in a positively biased estimation (Nickell, 1981). The fact that OLS is expected to provide a positive bias in the estimator if the idiosyncratic component exists is a useful check on the validity of the proposed cohorts studied. To draw out the fixed effect of the error term, a possible solution would be to apply a mean deviation transformation to equation (1). Under this transformation, the equation variables are expressed as a deviation from their period mean, thus eliminating the time-. 30.

(44) invariant cohort fixed effect and any other fixed variable (within estimation). But in panels with a short time span, the transformed autoregressive term (!-,% L # !-,% . . M%. ,!-, & N & !-,M .) is. now negatively correlated with the transformed error term (?0, L #. ?0, . . M%. ,?0 , & N & ?0,M .). as. the. !-,%. term. correlates. negatively with the M% ?0,% term in the transformed error. This . bias decreases as the time frame T becomes larger; hence, the. within estimators are asymptotically consistent on the time dimension. Bond (2002) points out that these different directions in the bias of both OLS and within estimators provide useful bounds on the accuracy of any other theoretical superior estimator proposed. Another approach that eliminates the unobserved fixed effects is to apply first differences to equation (1) as shown next:. O!-, # $ O!-,% & $ O,/0,% !-,% .. ---- ,., & $1 O/0, &$2 O,3 !-,% . & $4 O5. & O?0,. Where. ,7.. O!-,% # !-,% !-,%. As first differences are applied, any time-invariant regressor is also eliminated and can no longer be estimated. It can easily be shown. that the !-,% component in the transformed autoregressive term is correlated with the ?0,% component in the transformed error term. 31.

(45) by the dynamic nature of the model. Following the method proposed Holtz-Eakin, Newey, and Rosen (1988) and continued in Arellano and Bond (1991), a generalized method of moments (Hansen 1982) approach is taken to account for the endogeneity presented in (7). Building on the first moment conditions given by (3), (4) and (5), a set of instruments are available to account for this problem. As the number of available instruments is quadratic in the time dimension of the panel, many problems can be encountered in finite samples. Roodman (2006) presents various methods available to account for the over identification problem derived from large instrument matrixes. Another possible estimation method would be the use of system GMM, presented in Blundell and Bond (1998).5 The advantage of system GMM is that it permits the estimation of time-invariant variable parameters by the inclusion of both level and difference instrument sets, but it also requires the use of large instrument matrices. Due to the instrument proliferation relative to the small sample size that would occur if the previous estimation method were used, the traditional difference GMM approach was preferred. But, as previously noted, this eliminates from the estimation any time-invariant regressor.. 1.5.. Results. 5. This method requires an additional assumption that the deviation of the first observation from the steady state is uncorrelated with the fixed effect.. 32.

(46) The results of the estimation of the conditional and unconditional income mobility models are presented. As previously explained, the models are also estimated by the use of OLS and within estimation to obtain reasonable bounds for the autoregressive parameter and to evaluate the results obtained by GMM.. a) The Unconditional Model The unconditional model was estimated using difference GMM procedures. As previously indicated, it is expected that a heteroskedastic component exists in the error term; hence, a robust correction in the variance and covariance matrix of the errors is applied. A total of 140 observations were used in the estimation of the model (since one period is lost due to the application of first differences), and a collapsed instrument matrix6 containing a total of two instruments was used.. Table 5. Unconditional Model Results Coefficients Within OLS GMM Estimation Estimation Estimation QR,S%T P Constant. 0.476379*** (0.05998). 0.922871*** (0.02993). 3.438388*** (0.38695). 0.559065*** (0.19341). 0.864752*** (0.08274) -. Source: Authors’ calculations. Note: *** Significant at 1%. 6. See Roodman (2006).. 33.

(47) It can be observed that the GMM estimator for the autoregressive term is inside the bounds given by the OLS and within estimators. The total unconditional convergence is estimated at 0.86, slightly below the 0.9 obtained for the rest of Latin America (Cuesta et al., 2011). As expected, the OLS estimator is larger than the one provided by GMM methods. This supports the belief that the cohorts are adequately constructed, as the unobserved cohort fixed effect is present, (hence the OLS estimation is positively biased). But due to the robust estimation method applied to the GMM estimation, its 95 percent confidence interval is wide and includes the value estimated by OLS; thus, the difference observed is not significant. Additional relevant statistics for GMM estimation are presented in Table 6.. Table 6. Unconditional Model Relevant Statistics Statistic. P-Value. Arellano-Bond Test. 2.95. 0.003. Sargan Test*. 0.03. 0.870. Hansen Test *. 0.05. 0.824. Source: Authors’ calculations Note: *The statistic has a chi-squared distribution with 1 degree of freedom.. 34.

(48) The Arellano-Bond test for autocorrelation is used to determine if the errors are serially uncorrelated (assumption (2)). The null hypothesis is that no second-order autocorrelation is present in the transformed error component (which translates to first order autocorrelation in level equation (1)); and as the hypothesis is not rejected, the assumption that there is no autocorrelation present in the error holds. The Sargan and Hansen tests are used to determine the quality of the instrument matrix used. The null hypothesis is that the instruments are not exogenous (which means that instrument matrix is not valid). Thus, for the consistency of the GMM estimators, this test must be rejected. The Sargan test is not robust, and the Hansen statistic is a robust measure but its validity is reduced as the number of instruments used increases (a frailty not shared by the Sargan statistic). Therefore, high p-values obtained for this test are generally construed as a warning of misspecification. However, taking into account the small number of instruments used and the fact that the Sargan test also presents high p-values, the hypothesis of exogeneity of the instruments is confidently rejected.. b) The Conditional-Entrepreneurship Model. ---- ,., vector: Only one cohort covariant was considered for the 5. the average number of residents whose income represents more than. 25 percent of their household income for cohort 7 in period (. This 35.

(49) measure is considered to be much more volatile in time than any other household covariate (including age of the household head, number of residents per household, education level for older cohorts, etc.), and thus would be able to survive the first differences taken. The following results were obtained:. !-,% /0,% !-,% 3 !-,% ---5 ,., /0,. Constant. Table 7. The Conditional Model: Results Coefficients OLS Within GMM Estimation Estimation Estimation 0.651593***. 0.7739***. (0.10039). (0.12009). -0.369551*** (0.10023). -0.16449* (0.08786). -0.24995* (0.12764). 0.002318 (0.00608). -0.27818** (0.10851). -0.297292* (0.15998). 0.054548 (0.09872). 1.423304*** (0.20309). 2.237625*** (0.44658). 3.087679*** (0.56992). 0.891305 (0.57498). 1.793539** (0.84013). 0.292977 (0.25073). 1.030288** (0.46863). -. 0.944132*** (0.0385). Source: Authors’ calculations. Note: * Significant at 10%;** Significant at 5%;*** Significant at 1%.. 36.

(50) As occurred in the unconditional model, the first autoregressive term is inside the bounds given by the other estimation methods. The difference between the OLS and GMM estimators is more notorious in the conditional model, but it is still not significant at a 95 percent level. This will be very difficult to accomplish considering the small size of the pseudo-panel and the robust estimations made. ;!-, # 0.7739 0.25/0,% 0.29733 ;!-,%. The total income mobility can be expressed as follows:. This means that an increase of 1 percent in the level of entrepreneurship in a cohort in period ( 1, translates into an. increase of 0.0025 in the income mobility of said cohort in period (.. An interesting result is that female cohorts experience significantly higher income mobility, as their base convergence level can be expressed as (since 3 # 1 for female cohorts):. ;!-, # 0.4766 0.25/0,% ;!-,%. To complement the reduction of income convergence that occurs with an increase in the percentage of entrepreneurs in a given cohort, this increase also positively affects future income. Additional relevant statistics for GMM estimation are presented below:. 37.

(51) Table 8. Conditional Model: Relevant Statistics Statistic. P-Value. Arellano-Bond Test. 1.59. 0.111. Sargan Test*. 33.4. 0.001. Hansen Test *. 17.15. 0.192. Source: Authors’ calculations. Note: *The statistic has a chi-squared distribution with 13 degrees of freedom.. The Arellano-Bond test is rejected at a 10 percent level, so firstorder correlation in the error term is to be expected. On account of this issue, assumptions (2), (3), (4) and (5) need to be corrected as =,?0, 5-,.,%A . # 0 B C D 1 =>?0, !-,% @ # 0 B ( D 3 =>?0, ?0,%A @ # 0 B C 0 =,/0,%A :-, . # 0 B C 1 ---- ,.,%A :-, @ # 0 B C D 1 >5 E>?0, !-,% @ # 0 B ( D 3. follows:. The following instrument matrix was constructed: For every ( D 4 •. • • • •. !-,%U B C D 3,. /0,%U !-,%U B 3 3 !-,% B 3. /0,%. ---5 ,.,%. C D 4,. CD5. 38.

(52) As was done with the unconditional model, the instrument matrix constructed was collapsed to reduce the number of instruments without loss of information. This resulted in an instrument matrix with a total of 18 instruments and 120 observations available for the estimation of the model. The Sargan statistic is not robust and, due to the heteroskedasticity of the error term (guaranteed by the pseudo-panel approach taken), the fact that the null is not rejected should not be alarming. The Hansen test is rejected at a 10 percent significance level, but the value is not high enough as to generate doubt (an empirical rule of thumb is to consider p-values approaching 0.3 or higher as suspicious) so no miss specification signs are present.. 1.6. Conclusion There have been few attempts to measure the determinants of income mobility in Latin America, due mainly to the limited information collected in those countries and the lack of panel data. This is especially true for Ecuador, where few such studies have been undertaken. But building on recently developed methods of estimation without the use of panel data and other efforts to reduce the dependence of consistent autoregressive estimators on a large time frame in a dynamic panel scheme, estimations of income mobility and its determinants were achieved for the Ecuadorian case.. 39.

(53) Through the construction of a pseudo-panel and the application of difference GMM estimation methods, unconditional and conditional income mobility models were estimated. As expected, and in accordance with other empirical studies, the unconditional model tends to underestimate true income mobility, as there are other factors not included in the equation that explain future income and are correlated to previous income (there is a missing variable problem and estimations are biased). The inclusion of other cohort covariates in the proposed model reduces the bias and permits the analysis of other determinants of income mobility and future income (these might be affected by public economic policy). One of those factors, and the one of particular interest for this chapter, is the percentage of entrepreneurs in the different cohorts. The results of the GMM estimations revealed that entrepreneurship not only reduces income convergence by 0.0025 but also increases future average cohort income by 1.79 percent per percentage-point increase in the cohort’s entrepreneurship rate. This means that entrepreneurship not only positively affects income generation on average, but also makes it easier to generate such an increase. Another interesting result is that households headed by women tend to experience more income mobility. This, paired with the fact that their income is significantly lower than those of male-headed households, is a clear indication of the vulnerability of femaleheaded households.. 40.

(54) It is also important to notice that the difference between the OLS and GMM estimators is more notorious in the conditional model. This supports the belief that the cohorts are adequately constructed; as the unobserved cohort fixed effect is present, (hence the OLS estimation is positively biased). Nevertheless, this difference is not significant at the 95% level, and this will be very difficult to accomplish considering the small size of the pseudo-panel and the robust estimations made. The classification of entrepreneurship used in this study suffers from data censorship, as only families currently in charge of a business are considered entrepreneurs. Households that owned a business but went bankrupt would fall in the non-entrepreneur category. This leads not only to an overestimation of the positive effect mentioned in future income but also to a possible further increase in the positive effect on income mobility. However, considering that under normal conditions most businesses that go bankrupt follow a process that takes time, the censorship should not be problematic. This is because in any given period, some entrepreneurs are thriving while others are failing. This effectively reduces the average income they perceive over time and considerably diminishes any possible bias. The results of this study indicate that public policy should put special emphasis on promoting incentives for the development of entrepreneurship as a strategy for economic development.. 41.

(55) Appendix Table 1. Distribution of the Households Sampled by Gender and Date of Birth Cohorts Years Analyzed Cohorts. Total. Men Women. 2003. 2004. 2005. 2006. 2007. 2008. 2009. 2010. 1. 1521. 1487. 1401. 1328. 1156. 1176. 1145. 1121. 10335. 2. 1399. 1601. 1325. 1392. 1205. 1259. 1315. 1405. 10901. 3. 1535. 1523. 1438. 1420. 1524. 1682. 1645. 1755. 12522. 4. 1549. 1510. 1518. 1501. 1357. 1308. 1388. 1426. 11557. 5. 1150. 1353. 1251. 1242. 1198. 1183. 1247. 1322. 9946. 6. 1491. 1412. 1337. 1445. 1200. 1385. 1382. 1406. 11058. 7. 1506. 1577. 1548. 1339. 1367. 1402. 1328. 1516. 11583. 8. 1655. 1731. 1595. 1803. 1728. 1723. 1725. 1774. 13734. 9. 1671. 1639. 1497. 1515. 1794. 1789. 1552. 1581. 13038. 10. 961. 1038. 1220. 1347. 1862. 1758. 1787. 1933. 11906. 11. 680. 657. 642. 620. 526. 612. 564. 589. 4890. 12. 505. 580. 530. 570. 500. 536. 506. 602. 4329. 13. 486. 495. 429. 482. 522. 582. 625. 727. 4348. 14. 434. 435. 381. 415. 417. 408. 459. 531. 3480. 15. 305. 360. 310. 353. 342. 342. 401. 453. 2866. 16. 356. 346. 314. 317. 349. 427. 417. 442. 2968. 17. 337. 356. 307. 319. 361. 386. 350. 435. 2851. 18. 312. 329. 309. 391. 366. 456. 439. 467. 3069. 19. 211. 252. 258. 267. 303. 350. 370. 432. 2443. 20. 195. 209. 225. 279. 332. 368. 406. 473. 2487. 18259. 18890. 17835. 18345. 18409. 19132. 19051. 20390. 150311. Total. Source: Calculated by Authors based on ENEMDU surveys.. 42.

(56) Appendix Table 2. Distribution of the Households Weighted by Gender and Date of Birth Cohorts Years Analyzed Cohorts. Total. Men Women. 2003. 2004. 2005. 2006. 2007. 2008. 2009. 2010. 1. 222185. 213682. 217222. 204647. 172527. 181443. 181781. 171545. 1565031. 2. 195668. 238828. 205626. 219123. 198276. 201554. 226031. 226984. 1712091. 3. 244078. 241450. 237137. 227344. 250774. 276392. 277756. 284957. 2039888. 4. 239114. 242831. 258934. 260577. 235487. 234598. 243217. 246474. 1961232. 5. 179716. 220244. 219791. 225065. 216272. 212070. 217332. 218750. 1709240. 6. 236965. 224501. 234392. 253949. 219800. 247841. 238696. 241102. 1897246. 7. 240246. 255779. 268285. 247245. 248280. 257454. 238871. 260530. 2016690. 8. 264702. 276059. 270947. 303896. 317966. 304104. 307826. 310869. 2356368. 9. 267446. 277079. 279140. 263016. 328642. 327426. 284393. 277867. 2305008. 10. 154527. 166747. 216726. 240924. 347137. 328959. 341834. 352987. 2149840. 11. 99501. 96330. 99457. 102256. 88582. 100515. 97725. 98631. 782996. 12. 76802. 86569. 87160. 92989. 84152. 89649. 88349. 105824. 711495. 13. 80186. 81541. 74399. 84685. 95134. 105482. 114474. 129457. 765357. 14. 72619. 74104. 72840. 81452. 82510. 78154. 81774. 92864. 636317. 15. 55332. 63672. 59513. 68102. 63566. 61199. 71520. 82187. 525090. 16. 58501. 59247. 54766. 57536. 70491. 89688. 78716. 84324. 553269. 17. 53132. 59559. 59351. 61261. 67628. 74120. 69591. 88904. 533544. 18. 52854. 50106. 60060. 80833. 68931. 85654. 85545. 92480. 576463. 19. 35325. 37220. 48858. 45945. 58482. 62280. 74424. 82875. 445407. 20. 30134. 33381. 42505. 47290. 63648. 70016. 82707. 93536. 463217. 2859031. 299893 0. 306710 9. 316813 5. 327828 4. 338859 7. 340256 0. 354314 4. 2570579 0. Total. Source: Calculated by Authors based on ENEMDU surveys.. 43.

(57) 2. EXTERNAL. ECONOMIES. OF. SCALE:. FOREIGN DIRECT INVESTMENT AND FIRM SIZE IN ECUADOR 2.1 Introduction and Literature Review There is a broad literature on the determinants of firm size (For a survey, see Bernard and Muller, 2000). Economies of scale, economies of scope or transaction costs are some of the most common explanations. Other papers such as Evans (1987) suggest the importance of growth and age in determining the size of firms. With the somewhat recent boost of entrepreneurship, the relevance of determining what makes a firm larger has gained importance.. On the other hand, although the role of FDI in economic growth is generally perceived as positive, the channels though which this effect is generated is not all clear. Foreign Direct Investment is perceived in many countries as a way for development. Some authors, such as Al-Sadig (2013) have found that FDI incentives domestic investment. The crowding in effect of FDI is noted in Göcer et al (2014) and other authors.. 44.

(58) Agglomeration externalities allow new catalysts. for gowth and. economic development.. In this chapter, we analyze the role of industry level foreign direct investment (sectorial investment) on the firm size of Ecuadorian firms. We pose it to be an external economy of scale. Following Broadberry and Marrison (2002), we focus on external economies of scale as a source of cost-benefit and thus enabling firm size increase. In particular, Broadberry and Marrison differ internal from external economies of scale:. “Internal economies of scale are a fall in unit costs arising from the expansion of an individual firm, and hence are not necessarily associated with an increase in the scale of the industry. External economies take the form of a fall in unit costs arising from an expansion of the industry without an increase in the size of individual firms”. Other authors that explore the role of external economies of scale include Chipman (1970) that formally defined the concept. However, in other contexts it was used long before (see for example 45.

(59) Hoover 1937 that used the concept to explain urbanization externalities).. Among the drivers of external economies of scale, authors like Ciccone (2002) or Wheeler (2001) have shown that they lead to better access to human capital decreasing also the search costs of labor force.. It is of particular interest to analyze the effect of Foreign Direct Investment in Ecuadorian firms, given the low levels of FDI received by Ecuador in the recent years: it is among the lowest of the region (World Bank, 2015) and in particular significantly lower than its neighbors Colombia and Peru. Also, being a country that has benefit from the oil boom (from 2007 to 2014),. The rest of the chapter is organized as follows: Next section presents the theoretical model, and the data implications. Then, we analyze the results of the dynamic estimation under three econometric scenarios. Finally we present some concluding remarks and future work.. 46.

(60) 2.2 The Model: Determinants of Firm size As discussed in the previous sections, the analysis of the dynamics in firm size has been approached in literature in several different ways. The present study seeks to evaluate the determinants of firm size and its development through time, distinguishing between industry level effects, referred to in literature as external economies of scale, and firm level determinants of scale and cost structure. Formally,. the size of a firm from an industry V in a given period ( can be represented as follows:. C # $W & XC% & $ & $ & $1 % & $2 Y & $4 Y% &. $6 Z & $[ \] & $^ \]% & _ (1) Where,. C : Is a proxy variable for the size of the business measured as the total assets owned the firm on period (.. : Represents the number of employees working for firm on. period (. Moreover, the existence of diminishing returns in labor, given the size of the firm in question in the last period (C% . and. the other determinants considered, could be contrasted by evaluating if $ is significantly below 0. The % term is included. considering the difficulties a firm faces when deciding to reduce its. number of employees from one year to the next, mainly due to existing job contracts and regulations. 47.

(61) Y : Represents the total sales of firm on period (. The lag of this. regressor is included recognizing that, in some industries, consumers tend to form long term relationships with the business offering them goods and services. Z : Is the age of firm on period (. Hence, the relevance of the. firm’s life cycle in its growth (for the analyzed time period) could. be evaluated by $6 coefficient.. \] : Represents the Net Foreign Direct Investment (FDI) received. by industry V on period (.. This covariate serves as a proxy for the external economies of scale created by foreign investment as it percolates through a given industry by increasing its capital and the access to modern technologies. These in turn can be utilized by firms other than the primary beneficiaries of the investment. Due to the lack of information of FDI in a resolution lower than an industry level, the existence of the external economies of scale can only be evaluated on average for the different industries that constitute the Ecuadorian economy. A lagged term of this covariate is included considering that this type of investment could take more than one year to manifest as external economies of scale to businesses other than those directly receiving the transfers.. 48.

(62) X: Represents the autoregressive coefficient, which in this context could be interpreted as the growth inertia of a given firm (or the. proportion of a firm’s size that is explained only by its size the period before). _ = A composite error term determined by the next equation. _ # λ & ?. Where λ is a time-invariant intrinsic firm characteristic which. cannot be observed, and ? is an error term which is assumed to be. serially uncorrelated. The importance of this assumption is discussed below. The specified functional form allows for the contrast of several different theories of the dynamics of firm growth by assessing which of its determinants were significant in the Ecuadorian business environment for the years studied. Furthermore, unbiased estimators of the relevant model parameters, required to test the different hypothesis, can be obtained through dynamic panel estimation methods like the two-step least squares estimation (Anderson-Hsiao 1981) or a more efficient estimator obtained within the Generalized Method of Moments (GMM) framework.. a) Estimation by Dynamic Panel Methods The dynamic nature of the model presented renders static panel estimation methods biased because the estimation centers on following the same firm over time. This introduces an unobserved. 49.

(63) fixed effect that is correlated with the autoregressive term (Nickell 1981). The following expression is obtained by replacing the composite error term with its determinants in (1): C # $W & XC% & $ & $ & $1 % & $2 Y & $4 Y% & $6 Z & $[ \] & $^ \]% & λ & ? ,2.. By expressing this equation on the t 1 period, one can show that. the unobserved component is correlated with the previous firm size, which in turn introduces a missing variable problem that would render standard estimation methods biased. However, given the assumed relationship between the error term and past realizations of the dependent variables, different consistent estimation methods can be derived. For the model in question, the following assumptions are made: •. The error term is believed to be serially uncorrelated,. •. The number of employees and the firm’s sales are believed. =,? ?% . # 0. ,3.. to be endogenous,. =,%A ? . # E,%A ? . # =,Y%A ? . # 0. 2. ,4.. B CD. This means that the error term is correlated with contemporaneous or future employment levels or sales. This assumption is made because of the plethora of other firm level factors, which are not. 50.

(64) accounted for in the model, and which affect simultaneously its size, number of employees or sales. •. The Net Foreign Direct Investment is assumed to be exogenous, but not strictly so. >\]%A ? @ # 0. B CD0. ,5.. This assumption is made considering that the FDI measure used is. at an industry level and that, in the context of a given firm , its contribution to this aggregate would be marginal. Hence, this covariate would be independent to any omitted firm characteristic. •. Building on the belief that the errors are serially uncorrelated, a natural supposition is made: E,C% ? . # 0. ,6.. Thus, as previously mentioned, the estimation of the proposed model without accounting for the unobserved cohort effect will result in biased estimators. Moreover, it can be shown that ordinary least squares (OLS) will result in a positively biased estimation. To draw out the fixed effect of the error term, a possible solution would be to apply a mean deviation transformation to equation (1). Under this transformation, the equation variables are expressed as a deviation from their period mean, thus eliminating the timeinvariant cohort fixed effect and any other fixed variable (within estimation).. 51.

(65) But in panels with a short time span, the transformed autoregressive term (C% L # C% . . M%. ,C% & N & CM .) is now negatively. correlated with the transformed error term (? L # ? . . M%. ,? &. N & ?M .) as the C% term correlates negatively with the. . ? M% %. term in the transformed error. This bias decreases as the. time frame T becomes larger; hence, the within estimators are asymptotically consistent on the time dimension. Bond (2002) points out that these different directions in the bias of both OLS and within estimators provide useful bounds to check the specification of the proposed model. Another approach that eliminates the unobserved fixed effects is to apply first differences to equation (2). It can easily be shown that. the C% term in the transformed autoregressive covariate is. correlated with the ?% component in the transformed error term.. Following the method proposed in Holtz-Eakin, Newey, and Rosen. (1988) and continued in Arellano and Bond (1991), a generalized method of moments (Hansen 1982) approach is taken to account for the endogeneity created in the differenced model. Building on the first moment conditions given by (4), (5) and (6), a set of instruments are available to account for this problem. As the number of available instruments is quadratic in the time dimension of the panel, an over-identification problem can be encountered in finite samples. Roodman (2006) presents various methods available to account for issues for this issue.. 52.

(66) One drawback of the aforementioned methodology is that, as first differences are applied, any time-invariant regressor is also eliminated and can no longer be estimated. This is also true for any covariate which varies slowly in time (or in a stepwise manner). Hence, the application of this technique would eliminate the proposed model’s capability to contrast several hypothesis of the determinants of a firm’s growth like: the importance of its life cycle (captured by its age) or the relevance of business scale measured by a slowly changing covariate (number of employees). To account for this issue and in order to increase efficiency, Blundell and Bond (1998) proposed that instead of estimating the differenced model with the instruments, derived from the moment conditions outlined before, the level equation should be estimated with the use of differenced instruments. For this to be valid, as Roodman (2009) outlines, it is required that any change in an. instrumenting variable a be uncorrelated with the unobserved fixed. effect: =,∆a λ )=0. This implies that =,a λ . is time-invariant.. Moreover, Blundell and Bond show that for this to hold, an assumption must be made about the initial conditions of the data generating process which is that, controlling for the covariates, faster-growing individuals are not systematically farther or closer to their steady states than slower-growing ones. This is a reasonable assumption in the context of the proposed model, as the introduced covariates account for differences in both the life cycle and scale of the different firms.. 53.

(67) In order to make use of the additional moment conditions given by the method outlined, without losing those obtained by the original Arellano-Bond estimator, Blundell and Bond estimated a stacked data set consisting of both the differenced and level information for each individual (System GMM). In the following section the results of implementing this estimation method to (2) are presented.. 2.4 Results The results for the estimation of the proposed model are presented next. It is important to note that the applied methodology requires several modeling choices regarding the shape of the instrument matrix and the estimation of efficient and robust variances. Moreover, the risk of over-fitting endogenous variables is high within this framework on account of instrument proliferation, an issue that, as previously discussed, would render biased coefficients. To account for this, a detailed explanation of the number of instruments used and the type of matrix constructed is presented. This information is complemented with the exposition of several tests of autocorrelation in the error term and overidentifying restrictions which serve as flags for specification errors. In addition, as previously explained, the models are also estimated by the use of OLS and within panel estimation to obtain reasonable bounds for the autoregressive parameter and to evaluate the results obtained from System GMM.. a) The Instrument Matrix. 54.

(68) Building on the moment conditions defined in (4), (5) and (6) a collapsed7 instrument matrix containing 17 instruments was constructed. The following instruments were considered: •. As the dependent variable is predetermined, lags 1 through 4 were considered as valid instruments. •. For the not strictly exogenous variable FDI, the contemporaneous realization and its 3 first lags were used.. •. The age of the firm was used as a standard instrument (it instruments itself), which means that it was added as one column of the instrument matrix. Due to the fact that the age increases by one each year, only the untransformed version of this instrument was considered.. •. For the endogenous variables, namely number of employees and sails, only lags 2 and 3 are used as instruments. It is important to note that, to avoid instrument proliferation, only one of the employee. •. measures ( ) was used as instrument.. The System GMM methodology also includes the differenced values of the instruments to correct for endogeneity in the level equation. Due to its nature, as previously mentioned, the age of the firm was not differenced and added to this set.. 7. Roodman (2009) shows the superiority of collapsed instruments with simulations in some common scenarios.. 55.

(69) b) Estimation Results The estimation results are presented in table 2.1. In order to obtain a robust estimation to heteroskedasticity or serial correlation in the errors, a two-step estimation is applied. However, Arellano and Bond (1991) show that this produces standard errors which are downward biased when the instrument count is high, making inference unreliable. To account for this issue a Windmeijer (2005) correction is applied to the variance of the two-step estimates.. Table 2.1. Determinants of Firm Size: Estimation Results. Coefficients. C,% \], \],% ,. OLS. Within. GMM. Estimation. Estimation. Estimation. 0.8987991*** 0.131757***. 0.686546***. (0.00114). (0.00371). (0.21209). 0.4085***. 0.504826***. 0.400271***. (0.13396). (0.13409). (0.11039). -0.47582***. 0.016748. 0.23323. (0.13283). (0.12853). (0.18066). 3717.707***. 4744.462***. 8762.21*. (97.24029). (148.5559). (5306.433). 56.